Repository navigation
Releases
Releases · noble1911/trace
trace v0.15.0
Your own boards
- A Boards page for boards you make in trace. New board… sets one up from a template (Software, Research, Bug triage or Blank) with its columns, card types, a kickoff prompt and the repositories its agents run in. Everything can be changed afterwards.
- Cards (
local-12) start agents and move between columns like any ticket. A card type can have its own prompt, so a Bug and a Spike on the same board start differently. - A board can have its own icon.
Board settings, on every board
- Any board, Jira's and Pylon's too, has ⋯ → Board settings…: its own kickoff prompt and its own repositories. With one repository trace never asks; with several it uses the default, or asks and offers only those.
- The brief an agent starts with is the card type's prompt, else the board's, else the global one. New placeholders:
{tracker},{type},{board}. - A ticket's agent is told which ticket it's working on every time it starts ("Jira ticket PM-1234"), including after a resume.
One Boards button in the sidebar
- The board pages are grouped under one Boards button. The first click opens the board you last used; click again to open the list beneath it, and again to close it.
- The list has Jira and Pylon once each, then each of your own boards. All boards lists every board by name. Settings → Sidebar turns the grouping off.
Connect Jira with your login
- Integrations → Jira has Connect with your Atlassian login: no API token needed. It uses Atlassian's MCP server (v2). An API token still works, and is used if one is saved.
- Reading boards this way was checked against a real Jira site. Moving a card and commenting haven't been run against one yet, so tell us if a move fails. Two differences from an API token: no linked pull requests, and "hide completed issues older than" follows Jira's default, not your board's setting.
- Pylon has the same option, but it hasn't been tried against Pylon's real server yet. Treat it as experimental.
MCP servers
- Atlassian moved to its v2 server; an existing login gets a Move to new server button.
- Pylon and Intercom have EU servers. With the Pylon tracker connected, only your region's is offered.
- A Metabase card: enter your instance's address and a name. Add one per instance (staging, production).
- A server already in your Claude config keeps its name instead of being added twice.
Agents can work with every board
- trace's tools for agents now cover all boards:
list_boards,get_boardfor any board with its columns and setup, and moving, commenting on and starting agents for cards on boards that aren't on screen. On your own boards they can also create and update cards.
Also new
- Scheduled prompts can run as a chat, not only in a terminal. Thanks Lembit (#101).
- Pull requests is always in the sidebar and lists pull requests from sessions and ticket workspaces, not only Jira-linked ones.
- trace keeps the computer awake while an agent is running (the display can still sleep). Settings → Agents turns it off.
- A finished subagent's tab closes by itself five minutes after it ends if you never opened it. Its card in the chat brings it back. Settings → Tabs turns it off.
- The assistant's model is picked from the models your Claude Code offers (Settings → Assistant).
- Import from Conductor is offered on the empty Sessions page and at the top of Settings → Repositories, and imported sessions show at once.
- The Fleet page says Jira ticket or Pylon ticket.
Fixes
- Buttons that show a busy label ("Start" / "Starting…") centre their text instead of leaving a gap on the right.
- The homepage's download buttons point at the newest release that has a file for your platform.
v0.14.1
New models show up by themselves
- The model pickers (the chat's model menu and the New session dialog) now list the models your installed Claude Code offers, under the names Claude Code uses. When Claude Code gets a new model, it appears in trace without a trace update. Fable 5.2 is the first.
- Fixed: Opus 5.5 was listed twice while it was running.
Slash commands you can see
- A / button beside the paperclip opens the full list of slash commands. Typing
/does the same. - Each command shows what it does and what it takes, e.g.
/compact [instructions]. The list scrolls instead of stopping at eight. - The commands are listed before the agent has started, too.
Name a session up front
- The New session dialog has a Name field under the task. Leave it empty and the session is still named from the task.
Small fixes
- Add your own… on Integrations → MCP servers no longer wraps onto two lines.
- Settings sections have more room between their buttons and the text below.
v0.14.0
Setup installs Claude Code and the GitHub CLI for you
- On a Mac or Linux computer that's missing them, the setup checks now have an Install button on the Claude Code and GitHub CLI rows. No terminal, Homebrew or admin password needed. The GitHub CLI download is checked against GitHub's published checksum before it's installed.
- Each also gets a Sign in button. Claude Code opens your browser and finishes by itself. For GitHub, trace copies the one-time code, opens GitHub's page and shows the code while it waits. The row turns green on its own when you're done.
- Set it all up does every install, then every sign-in, one after another.
- The setup checks now also notice when Claude Code is installed but not signed in, instead of letting your first agent stop at a login screen.
- Find them in Settings → Run setup again, or from the setup notice at launch. Windows keeps its install commands for now.
Start a session from the task
- The New session dialog now opens on What should the agent do? That task becomes the agent's first message, and the session is named from it.
- The dialog remembers the repo, agent, model and mode you last used; everything else is under More options. (Thanks @lemps.)
A permission mode trace reviews
- New Reviewed by trace mode: reads and trace's own tools go ahead without asking, and everything else comes to you as an approval card. Always allow keeps that rule for the rest of the run.
- Terminal agents get the same built-in rules, so
git status,git diffandgit logstop prompting. (Thanks @lemps.)
v0.13.5
Fixed: an agent's replies not showing
- A chat agent could answer and trace would show nothing: the reply was saved, but the conversation stayed blank, and asking again did the same. Replies now always appear, including in conversations where this had already happened. (Thanks @lemps.)
Ticket agents can run in a Vercel sandbox
- The Start menu on a ticket has a Runs on choice, This machine or Vercel sandbox, shown while Vercel is connected. Sessions could already do this; tickets can now too.
- Settings → Agents has a matching default, used when a card auto-starts and when the orchestrator starts an agent. If a sandbox can't be started, the agent starts on this machine and says so.
- A sandbox ticket runs Claude, and only repositories with a GitHub remote can be picked.
- Not available on a sandbox ticket yet: the Terminal and Tests tabs, Raise PR, Merge and the open-in-editor menu.
(Thanks @lemps.)
v0.13.4
Subagents are tabs
- Each subagent has its own tab beside its agent, in the same tab bar as Files and Terminal. The second row of tabs inside the chat is gone.
- The bar under the tabs follows the tab you're on: the model, cost and context for an agent; "Claude › subagent name" with its state, time and calls for a subagent.
- A subagent's tab is for reading. The message box lives on the agent's tab. Esc on a subagent's tab takes you back to its agent.
- Tabs give up their labels one at a time when the window is narrow, subagents first. Before, they all collapsed at once.
- The chat remembers where you were reading when you switch to another tab and back.
- A tab's × no longer vanishes as you move the pointer onto it.
A simpler composer
- Model and effort are one control: "Opus 5.5 · High". The permission mode has its own glyph.
- Model ids read as names ("Opus 5.5", not "claude-opus-5-5").
- In a narrow pane the pickers shrink to icons instead of being cut off.
- With a provider that fixes the model, you can still change the effort.
Questions and approvals
- One card everywhere: the same neutral card with an amber border in the chat, the Inbox and the terminal view.
- Option previews show. When an agent attaches a code preview to each option, it now appears beside the options. Before, trace dropped them and the agent's question referred to previews you couldn't see.
- No more "Working" under a question. While an agent waits for your answer the chat says so.
- Long option descriptions wrap, and "(Recommended)" is a small tag.
- Inbox: the last message is formatted, the reply box has a Send button with "Send and go to next", and the answer is the highlighted button while there is one.
Settings
- Changing a setting no longer moves the page. Density and Interface size keep the control under your pointer, typed values apply on Enter, previews stay a fixed size, and settings that don't apply are dimmed with the reason.
- Appearance
- System can use your choice of light theme (Light or Paper) and dark theme (Dark or Midnight).
- Zoom with ⌘=, ⌘− and ⌘0; Interface size now runs from 80% to 150%.
- A code font for the whole app: ids, diffs and the Files tab follow it.
- Terminal scrollback is 10,000 lines by default (was 1,000) and can go up to 100,000.
- Repositories is one aligned list. Each row shows its remote, default branch and setup, with a ⋯ menu for the rest. A repository whose folder is missing says so.
- Worktrees are grouped by repository and named by their session or ticket, not by id. What auto-cleanup will delete is marked on the list.
- Tabs are regrouped: Prompts (kickoff and saved prompts), Layout (sidebar, tabs, buddy) and About. Tabs that don't fit move into a menu.
- "Reset all" shortcuts and "Reset to default" now ask first; saving a key shows "Saving…" then "Saved".
Also
v0.13.3
PDFs that match what you see
- Save as PDF follows the Plan tab's proportions. It uses your Chat & transcript font, heading scale and line height at a paper size, under a one-line header. Before, the type was about a quarter larger against its column and the title was printed twice.
- HTML documents break across pages instead of being cut through the middle of a card. Cards, images and table rows stay whole, and collapsed sections print open.
- Wide tables fit the page, and copy buttons no longer print.
Deletes ask first
- Deleting an archived session, removing a worktree (one or all merged), cleaning up workspaces, removing a saved key, deleting a Sessions tab or section, and removing a repo mapping all confirm in place. Removing a worktree with uncommitted changes warns there too, with no system dialog.
Menus everywhere, and the same everywhere
- Session rows and board cards have a ⋯ with everything right-click offers. Fleet and Inbox rows have menus for the first time.
- Every menu takes arrow keys, a menu's button looks pressed while it's open, and an item that's unavailable says why on a second line.
- The board switcher is a menu, and the terminal's right-click menu stays inside the window.
One chat design
- The orchestrator's chat looks and works like an agent's chat: the same message bubble, send button and confirm cards, whose button now names the action (Move, Start, Send) instead of "Confirm". It follows Settings → Appearance → Chat & transcript.
- Transcript and Chat agree. The same tool call has the same name in both ("Run", "Search"), edits show as diff cards and task lists as task cards, and the Working line has a timer.
- The composer's row is one height, and what you type is the size it's sent at.
- Cards that need you say what they ask: "Permission · Bash", "Question", "Plan approval".
Smaller things
- Small buttons inside rows are a proper size instead of full height; filter chips have the same corners as other controls.
- Buttons keep their width while busy ("Joining…", "Checking…").
- Errors from Start, Raise PR, Merge, opening a terminal, the repo picker and Settings arrive as toasts with the reason.
- A scheduled prompt's Edit and Delete are under its ⋯, and Esc goes back from its detail.
- Amber is used only for an agent that needs you: the usage bar and the Stats flags are neutral.
- Enter on a button inside a row no longer also opens the row.
- Labels end in "…" only when they open a dialog or picker.
v0.13.2
Steer a chat agent while it works
- Enter during a turn steers it: the agent reads your message after its current step, instead of waiting for the turn to end.
- ⌥↩ queues the message for when the turn ends.
- ⌘↩ (Ctrl+Enter on Windows/Linux) interrupts and sends: stops the turn, then sends your message first.
- Once you type while the agent works, the send button gains a caret listing these three. A queued message can be sent into the running turn with Send now.
The new shortcuts can be changed in Settings → Shortcuts.
Chats show their whole conversation
- A session that ran in the terminal no longer opens empty as a chat: its terminal turns show in the chat, in order, before you press Start.
- Turns taken in the terminal between two chat runs are filled in too.
v0.13.0
Give your agents your MCP servers, here and on Vercel, and Vercel sessions that keep their secrets out of the sandbox and look after themselves.
MCP servers
- Integrations has a second tab, MCP servers: a card for each popular server (Sentry, Linear, Notion, Atlassian, GitHub, Figma, Vercel, Supabase, Neon, PostHog, Pylon and more), the HTTP servers in your Claude config, and any you add with Add your own…
- Connect once: trace opens the server's login in your browser (or takes a key, or needs nothing). trace keeps the login; agents never see it
- Each server has two switches: On this computer (agents reach it through trace, which adds a fresh login to every call, so long sessions never lose it) and In Vercel sandboxes
- A server your Claude config already gives local agents stays theirs unless you switch it on here
Sessions on Vercel
- No secret in the sandbox: your GitHub, Claude and model-provider logins are added by Vercel's network outside the VM. The agent works as before but can't read them
- The repo's setup commands run in the sandbox after the clone (files the setup copies, like
.env, stay on your computer) - A terminal on Vercel knows when a turn ends or Claude needs you, as one here does, and catches up on what happened while you were away
- The Pull request tab finds the PR the agent opened, whatever it named its branch
- The Files tab works on Vercel
- A chat on Vercel gets trace's own tools (publish a document, open a page, start a session…)
- Previews: ports 3000, 5173 and 8080 get public URLs, listed in the session's Details.
localhost:<port>in the Browser tab opens the sandbox's server - On Pro a sandbox keeps itself running while its agent works, with trace open or closed, and stops (and saves) once idle. Integrations → Vercel → Unattended sandboxes sets how long it waits idle (10 minutes) and the longest run (6 hours)
Also
- A chat on Vercel no longer stops at the sandbox's first time limit while you're using it
v0.12.4
The recent sessions list in the rail is now opt-in, and a scheduled run waiting on you stops flickering.
Sidebar
- Resting on Sessions in the rail no longer slides out your recent sessions unless you want it to. Turn it on in Settings → Sidebar → Recent sessions, where its delay setting now lives too
Scheduled prompts
- While a run waits on a question, the "Claude is waiting on you" note shows once instead of blinking on and off, so the terminal under it no longer flickers. The run's "needs input" badge and the rail's Scheduled badge hold steady too
v0.12.3
Copying from the chat works every time again and keeps its formatting. A chat that hasn't started now starts when you message it, and Markdown files preview properly.
Copy
- Copying no longer fails with "Couldn't copy to the clipboard". Every copy in trace now goes through the system clipboard
- A message's Copy and Copy selection keep their formatting: headings, lists, bold, code and tables paste formatted into Docs, Notes, Mail and Slack. Editors and terminals still get the Markdown
- Copy table pastes as a real table, with borders, into docs and as cells into spreadsheets
Chat
- Send a message to an agent that isn't running and it starts. The message waits in the queue until the agent is up, and stays there if the start fails
Files
- A Markdown file's Preview looks like a Plan document: the chat's text styles and clickable file links, on a centred reading column